Notes from the Intersection No. 002
Carmen Onchain · 28 May 2026 · 8 min read
Sovereign AI cannot run on someone else's servers
In a single week, three very different institutions said almost the same thing in three very different vocabularies. Canadian Prime Minister Mark Carney confirmed that Canada will release a national AI strategy next week, explicitly framing the work as a sovereignty question. Pope Leo XIV released a manifesto calling for robust regulation of AI, lifting the conversation out of policy seminars and into the moral vocabulary used by two billion people. And a Bangladesh-based outlet, Blitz, ran a headline that summarized the gap better than any white paper this year: AI sovereignty requires more than powerful models.
Three institutions, three registers, one underlying claim. The country that wants to control its AI cannot get there by writing rules alone, and cannot get there by training a model alone. The hard part is everything underneath the model. And the choices being made right now, while everyone is distracted by benchmark scores and parameter counts, will set the defaults for the next decade of how AI actually gets owned, governed, and shut off.
The model is the easy part. A national lab can train a credible model in a year. Writing AI legislation, even good AI legislation, takes about the same. Owning the compute, the storage, the identity layer, and the payment rails that the model actually runs on top of takes a decade and a different kind of political imagination. That is the sovereignty problem nobody has solved yet, and it is the one this week's announcements quietly circle without naming.
The sovereignty question is not which model your country trains. It is who can switch off the server that the model runs on.
Deep Dive
What sovereignty actually requires
What is actually happening here
Sovereignty is the word of the moment in AI policy, and the word is being used in at least three different ways at once. The first definition is regulatory sovereignty: a country writes the rules under which AI must operate within its borders. The EU AI Act is the cleanest example. The second is data sovereignty: the country requires that certain categories of data stay inside its borders or under domestic control. India's draft DPDP rules and the Philippines' new AI roadmap both push in this direction. The third is the one almost no national strategy has clearly defined: infrastructure sovereignty. Which compute, which storage, and which model weights are physically and legally controllable by the sovereign in question.
Regulatory and data sovereignty are achievable through law. Infrastructure sovereignty requires building. And the gap between the first two and the third is where most of this year's national AI strategies are going to quietly fall down.
Why this matters beyond the money
The instinct in most policy conversations is to treat sovereignty as a procurement problem. Buy more domestic chips. Subsidize a national champion lab. Train a model in the local language. These are reasonable steps. They are also insufficient, because they all assume that the underlying stack will continue to look the way the current generation of hyperscale clouds taught us to expect: a handful of very large providers, in a handful of jurisdictions, running infrastructure that the rest of the world rents.
That assumption is what blockchain quietly disrupts. A decentralized inference layer, in the most boring possible description, is a way to run AI workloads on hardware that nobody has the unilateral power to switch off. A decentralized identity layer is a way to log who used what model without handing the logs to a single platform. A decentralized payment rail is a way for an autonomous agent to transact without going through a closed fintech CLI. None of these are mature. All of them are real, shipping in production, and improving fast enough that they will be at parity with closed alternatives on several axes within twenty-four months.
The reason this matters beyond the obvious money is that the cultural conversation about AI has just shifted. Pope Leo XIV's manifesto moved AI governance from a Silicon Valley debate into a mainstream moral one. Two billion people now have a religious leader on record saying that the question of whose values shape AI is not optional. That changes which policy proposals are politically possible. It also changes which architecture choices look reasonable to a regulator who is no longer just listening to vendors.
What builders and operators should be watching
If you are building anything AI-adjacent in 2026, here is a practical exercise worth running on your own stack. Take the architecture as it stands today and write down, for each layer, the answer to a single question: who can shut this off without my permission?
Inference compute. Data storage. Model weights. Identity and authentication. Payment rails. Logging and audit trail. Run the list. Most teams, if they are honest, can name a single vendor on three or four of those lines, and most of those vendors are headquartered in one of two jurisdictions. That is not architecture. That is a relationship. And relationships shift when policy shifts.
The fix is not to rip everything out tomorrow. The fix is to identify the two or three layers where dependency is highest, watch the decentralized alternatives in those layers carefully, and run small experiments now so you have a real migration path when you need one. Decentralized GPU marketplaces like Akash and Render are not yet a one-for-one swap for AWS, but they are real and improving. Filecoin and Arweave are functional storage layers. Ethereum-based identity standards are quietly being adopted by enterprises that two years ago would not have considered them. The point is not to bet the company on any single piece. The point is to stop treating the closed stack as the only stack.
Canada becomes one of the first G7 countries to put an AI strategy on the table in a year when "sovereignty" has gone from regulator jargon to mainstream language. The substance will tell us whether Ottawa understands sovereignty as a procurement question or as an infrastructure one. If the strategy funds domestic compute and identity rails as seriously as it funds research labs, that is a signal worth watching. If it stops at industrial policy and rule-making, the gap between what Canada wants to control and what Canada actually controls will be wider next year than this one.
Squamish Chief · 28 May 2026
A 15.6X speedup on long-context inference is not a benchmark story, it is an economics story, because it pushes million-token agent workloads from financially absurd into routine inside the next funding cycle.
VentureBeat · 28 May 2026
This is the cleanest single-sentence summary of where national AI policy is about to fail, and the clearest opening yet for decentralized inference and storage layers to argue that they are the missing half of the answer.
Blitz · 28 May 2026
When the Vatican puts AI governance into a moral framework that two billion people use to think about right and wrong, the political weather around regulation changes faster than most founders are pricing in.
NBC 6 South Florida · 26 May 2026
The first big closed-fintech move to court autonomous agents directly, which means the rails the agent economy actually uses are being set right now, by someone, somewhere, with or without an open-source answer.
Cxotoday · 28 May 2026
This is a week where regulators, religious leaders, and builders ended up in a strange and useful kind of alignment. They all said, in different vocabularies, that AI cannot be allowed to run on infrastructure that the people it affects do not control. That message has never sounded less like a fringe position.
The encouraging part is that the technology to actually act on that idea is finally arriving in the same season as the political will to demand it. Decentralized compute, identity, and payment rails are not theoretical anymore. They are early, but they are real. And early is the right time to be paying attention, because early is when the choices that compound for a decade are quietly made.
What is AI sovereignty?
AI sovereignty is the ability of a country, organization, or individual to control the models, data, and infrastructure used in AI systems that affect them. It typically combines regulatory rules, data residency requirements, and choices about which underlying compute and storage the AI actually runs on.
Why does decentralized infrastructure matter for AI sovereignty?
Decentralized infrastructure matters because regulatory and data sovereignty mean little if the underlying compute and storage are owned by a small number of foreign providers who can change pricing, terms, or access. Decentralized inference, storage, and identity layers reduce single points of control so that sovereignty exists in the architecture itself, not just on paper.
What should founders do about AI sovereignty in 2026?
Founders should map their current AI stack layer by layer and identify which providers could unilaterally shut off any part of it. Then they should run small experiments with decentralized alternatives in the highest-risk layers, like compute or storage, so they have a real migration path before policy or pricing forces one. This guide on blockchain and AI convergence is a good starting point for the architecture choices involved.
Who governs AI when sovereignty is decentralized?
Decentralized AI sovereignty does not mean ungoverned AI. It means governance is distributed across protocols, communities, and standards rather than concentrated in a few large companies. This guide on who governs AI walks through how DAOs, open standards, and audit trails can replace platform control without giving up accountability.